AI Paper Atlas

Research map / Label-efficient and robust learning / Self-supervised and contrastive

ZeoBench: A Benchmark for Self-Supervised Learning on 3D Zeolite Representations

Aaron Sun, Yachan Liu, Ping Yang, Peng Bai, Subhransu Maji

NeurIPS 2026 · Label-efficient and robust learning › Self-supervised and contrastive · approach: Invariance-based self-supervised learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ZeoBench, a benchmark for evaluating self-supervised learning on 3D zeolite crystal structure representations, motivated by scarce property labels for zeolite structure-molecule pairs.

Problem

Identifying zeolite materials optimized for a target application (gas storage, separations, catalysis) is hard because property labels are expensive and available only for limited structure-molecule pairs, despite hundreds of thousands of known zeolite structures.

Approach

A benchmark for self-supervised learning on 3D zeolite representations, aimed at leveraging large unlabeled zeolite structure collections to reduce reliance on scarce property labels. Details beyond the abstract are not available.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
in-domain SSL pretraining before fine-tuning instead of training from scratch3D ResNet-18 · ZeoBench C18, 10k labelslog(kH) MSE: 6.2 -> 4.9 (SimCLR)Improved
pretraining benefit grows as labels decrease3D ResNet-18 · ZeoBench fine-tuning, varying subset sizeslog(kH) MSEImproved
use contrastive objectives (MoCo, SimCLR) instead of BYOL or DINO for frozen features3D ResNet-18 linear probe · ZeoBench linear probe, 10k labelslog(kH) MSE: 27.4 (MoCo) vs 34.3 (BYOL) vs 44.1 (DINO)Improved
fine-tune after pretraining: BYOL versus contrastive methods3D ResNet-18 · ZeoBench fine-tuning, 10k labelslog(kH) MSE: BYOL 12.4 vs MoCo 12.6No clear effect
DINO pretraining before fine-tuning3D ResNet-18 · ZeoBench fine-tuning, 10k labelslog(kH) MSE: 13.5 -> 13.4No clear effect
add scale augmentation (sigma > 0) during SSL pretraining (ablation)SSL pretraining · ZeoBench C18 10k linear probelog(kH) MSEImproved
vary scale augmentation strength sigma (ablation)SSL pretraining · ZeoBench C18 10k linear probelog(kH) MSENo clear effect
use CNN (ResNet) instead of ViT or Swin trained from scratch (ablation)3D architectures · ZeoBench fine-tuning, 10k labelslog(kH) MSE: ViT 40.4, Swin 38.4 -> ResNet 13.5Improved
adapt ImageNet 2D ViT to 3D via depth tokenization (ablation)linear probe baselines · ZeoBench linear probe, 10k labelslog(kH) MSE: 56.9 (handcrafted) / 60.8 (3DINO) -> 48.7Improved
use test-time augmentation averaging for ImageNet adaptation (ablation)ImageNet depth tokenization · ZeoBench linear probelog(kH) MSEImproved
handcrafted descriptors in low-label regimeshandcrafted MLP baseline · ZeoBench linear probe, low labelslog(kH) MSEHurt
transfer ZeoBench SimCLR pretraining to hMOF adsorptionResNet-18 · hMOF, 5,000 training MOFsaverage MAE (mol/kg): 0.375 -> 0.242Improved

Experimental setup

Benchmarks: ZeoBench

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy